[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Anyone tried cupy?



https://salsa.debian.org/science-team/cupy
I had some attempts to package it with CUDA. But I forgot why I did not finish it and whether I was stuck somewhere.

I use numpy extensively, but not cupy. Since numpy is very important, and cupy can somehow work as a drop-in alternative, I think it would be good to investigate. Afterall with only PyTorch being the only ROCm validator is not ideal.

That said, if manpower is a concern, I would rank pytorch ahead of cupy.

On 6/26/24 05:14, Cordell Bloor wrote:
Hi folks,

I noticed that all the required dependencies for cupy with AMD GPU support are available on unstable [1]. I've never used cupy, so I probably won't file an RFS. However, it seems like a useful library. It describes itself thusly:

> CuPy is a NumPy/SciPy-compatible array library for GPU-accelerated computing with Python. CuPy acts as a drop-in replacement to run existing NumPy/SciPy code on NVIDIA CUDA or AMD ROCm platforms.

I have extended the architecture support of the AMD GPU libraries that cupy relies on, so it can probably be made to work with all discrete AMD GPUs from Vega onwards (though MI300 has not yet been enabled in Debian). With that said, the performance will undoubtedly be better on the architectures supported upstream.

Has anyone used cupy? If so, what was your experience?

Sincerely,
Cory Bloor

[1]: https://docs.cupy.dev/en/v13.2.0/install.html#using-cupy-on-amd-gpu-experimental


Reply to: